Dose Reduction in Step-and-Shoot 4D CT Imaging via Breathing Signal-Guided Deep Learning-Driven Selective Data Acquisition: A Feasibility Study

Respiration-correlated computed tomography (4D CT) is essential for radiotherapy of abdominal and thoracic tumors. Current protocols mitigate motion artifacts by continuously acquiring projections throughout entire breathing cycles at various couch positions within the field of view. However, many of the acquired projections are not used for subsequent image reconstruction, leading to additional radiation exposure. In this study, we propose the first deep learning (DL)-based approach to selectively guide projection data acquisition during 4D CT in real-time, based on online analysis of the patient's breathing signal, aiming to acquire projection data only when relevant for image reconstruction. We train a long short-term memory (LSTM) network to predict suitable beam-on events for data acquisition. The feasibility study is based on 1,659 breathing signals of 369 patients (60/20/20 train/validation/test split on patient level). For the test set, the model achieves a reduction of the beam-on time and the imaging dose of 30% (median value; IQR: 22 to 37%), translating into 8–16 mGy dose reduction for standard 4D CT CT-DIvol values. Compared to an ideal setting, 98% of the predicted beam-on starting points align closely with corresponding ground truth time points. These results demonstrate the feasibility of respiratory signal-guided, DL-driven selective projection data acquisition, which promises substantial dose reduction with only minor impact on image quality.